Evaluation of echosounder data preparation strategies for modern machine learning models

Fish stock assessment and management requires accurate estimates of fish abundance, which are typically derived from echosounder observations using acoustic target classification (ATC). Skilled operators regularly assisted in classifying targets by software there has been an increasing interest toward machine learning to create improved tools. Recent studies have applied deep approaches data, however, algorithm data-preparation strategies (influencing model output) presently poorly understood standardization is needed enable collaborative research management. For example, a common pre-processing technique resample backscatter data coming measurements the original resolution coarser horizontal (time) vertical (range) directions. Using values volume backscattering coefficient obtained during Norwegian sandeel survey, we investigate resampling resolutions suitable for ATC convolutional neural network trained classify single data. This process known as pixel-level semantic segmentation. Our results indicate that it possible downsample if important information related characteristics not smoothed out. We also show performance when providing with contextual relating range. These findings will provide input fisheries standards contribute on-going development automated methods.